Build pytorch 2.3.1

This (properly) builds pytorch 2.3.1, including mkl support.

platform/Dockerfiles/pytorch/Dockerfile

@@ -0,0 +1,209 @@
+FROM debian:bookworm AS bookworm_cuda
+
+###
+#
+# Define environment
+
+WORKDIR /workspace
+
+
+###
+#
+# Set global environment variables
+
+ENV PYTORCH_VERSION="v2.3.1"
+ENV CUDA_VERSION="12.5.1_555.42.06"
+ENV PATH="$PATH:/usr/local/cuda/bin"
+ENV DEBIAN_FRONTEND="noninteractive"
+ENV NVIDIA_DRIVER_CAPABILITIES="compute,utility"
+ENV NVIDIA_VISIBLE_DEVICES="all"
+ENV VENV_PATH="/workspace/v"
+ENV PYTHON_VENV="${VENV_PATH}/bin/python"
+ENV PIP_BIN="${VENV_PATH}/bin/pip"
+
+###
+#
+# Workaround gcc-12 issue:
+# https://github.com/pytorch/pytorch/issues/77939#issuecomment-1526844015
+
+ENV CXXFLAGS='-Wno-maybe-uninitialized -Wno-uninitialized -Wno-free-nonheap-object -Wno-dev'
+ENV CFLAGS='-Wno-maybe-uninitialized -Wno-uninitialized -Wno-free-nonheap-object -Wno-dev'
+
+
+###
+#
+# Set pytorch specific build environment variables
+
+ENV REL_WITH_DEB_INFO="ON"
+ENV MAX_JOBS="32"
+ENV USE_CUDA="ON"
+ENV USE_CUDNN=1
+ENV USE_CUSPARSELT=1
+ENV USE_FBGEMM="ON"
+ENV USE_KINETO="ON"
+ENV USE_NUMPY="ON"
+ENV USE_NNPACK="ON"
+ENV USE_DISTRIBUTED="ON"
+ENV USE_TENSORPIPE="ON"
+ENV USE_GLOO="ON"
+ENV USE_MPI="ON"
+ENV USE_SYSTEM_NCCL="OFF"
+ENV USE_OPENMP="ON"
+ENV USE_FLASH_ATTENTION="ON"
+ENV USE_MEM_EFF_ATTENTION="ON"
+ENV PYTORCH_BUILD_VERSION="2.3.1"
+ENV PYTORCH_BUILD_NUMBER="1"
+ENV TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9;9.0"
+ENV CUDA_PATH="/usr/local/cuda"
+ENV CUDA_HOME="/usr/local/cuda"
+ENV CUDA_TOOLKIT_ROOT_DIR="/usr/local/cuda"
+ENV CUDA_NVCC_EXECUTABLE="/usr/local/cuda/bin/nvcc"
+ENV CUDA_INCLUDE_DIRS="/usr/local/cuda/include"
+ENV CUSPARSELT_LIBRARY_PATH="/usr/lib/x86_64-linux-gnu"
+ENV CUSPARSE_INCLUDE_PATH="/usr/include/x86_64-linux-gnu"
+ENV CUDNN_LIBRARY_PATH="/usr/lib/x86_64-linux-gnu"
+ENV CUDNN_INCLUDE_PATH="/usr/include/x86_64-linux-gnu"
+
+#ENV USE_UCC="ON"
+ENV USE_MIMALLOC="ON"
+ENV USE_NCCL="ON"
+
+#ENV ATEN_THREADING="NATIVE"
+#ENV USE_SYSTEM_LIBS ON
+
+###
+#
+# Install toolchain and system dependencies
+
+RUN apt update
+RUN apt -y install build-essential
+RUN apt -y install ca-certificates
+#RUN apt -y install ccache
+RUN apt -y install python3
+RUN apt -y install python3-full
+RUN apt -y install python3-venv
+RUN apt -y install python3-pip
+RUN apt -y install swig
+RUN apt -y install ninja-build
+RUN apt -y install git
+RUN apt -y install cmake
+RUN apt -y install gpg
+RUN apt -y install curl
+RUN apt -y install zstd
+
+RUN apt -y install libnuma-dev
+RUN apt -y install libssl-dev
+RUN apt -y install libzstd-dev
+RUN apt -y install libucx-dev
+RUN apt -y install libmpfr-dev
+RUN apt -y install libgmp3-dev
+RUN apt -y install libfftw3-dev
+#RUN apt -y install libmagma-dev
+
+
+###
+#
+# Not sure if or why these are needed
+
+RUN apt -y install libjpeg-dev
+RUN apt -y install libpng-dev
+
+#RUN /usr/sbin/update-ccache-symlinks
+RUN mkdir -p /opt/ccache
+#RUN ccache --set-config=cache_dir=/opt/ccache
+
+
+###
+#
+# Setup build environment and clone pytorch
+
+RUN mkdir -p /workspace/build
+RUN mkdir -p /workspace/${PYTORCH_VERSION}
+RUN mkdir -p /workspace/tmp
+RUN mkdir -p /workspace/added
+RUN mkdir -p /workspace/uncompressed
+RUN mkdir -p /workspace/target
+RUN mkdir -p /workspace/patches
+
+RUN git clone --depth 1 --jobs ${MAX_JOBS} "https://github.com/pytorch/pytorch" --branch "${PYTORCH_VERSION}" --recurse-submodules --shallow-submodules build
+
+###
+#
+#
+
+COPY /workspace/patches/pytorch-compute-86-override.patch /workspace/patches
+RUN patch --directory build -p1 pytorch-compute-86-override.patch
+
+###
+#
+# Install NVIDIA CUDA SDK
+
+RUN curl -LO https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64/cuda-keyring_1.1-1_all.deb
+RUN dpkg -i cuda-keyring_1.1-1_all.deb
+RUN apt-get update
+RUN apt -y install software-properties-common
+RUN add-apt-repository contrib
+RUN apt-get update
+RUN apt -y install cuda-toolkit-12-5
+RUN apt -y install libcusparselt-dev
+RUN apt -y install cudnn
+
+
+###
+#
+# Install Intel MKL BLAS
+
+RUN curl --location "https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB" | gpg --dearmor > /usr/share/keyrings/oneapi-archive-keyring.gpg
+RUN echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" > /etc/apt/sources.list.d/oneAPI.list
+
+RUN apt update
+RUN apt install -y intel-oneapi-mkl
+RUN apt install -y intel-oneapi-mkl-devel
+
+ENV MKL_VERSION="2024.2"
+ENV MKL_ROOT="/opt/intel/oneapi/mkl/${MKL_VERSION}/lib/intel64"
+ENV MKL_MODEL="ilp64"
+ENV MKL_LIBRARIES="-Wl,--start-group;${MKL_ROOT}/libmkl_intel_${MKL_MODEL}.a;${MKL_ROOT}/libmkl_gnu_thread.a;${MKL_ROOT}/libmkl_core.a;-Wl,--end-group"
+ENV CUDA_ARCHS="80;86;89;90"
+ENV BLA_VENDOR=Intel10_64ilp
+ENV BLA_STATIS=True
+
+
+###
+#
+# Install Python virtual environmnet
+
+RUN python3 -m venv ${VENV_PATH}
+RUN ${PIP_BIN} install six
+RUN ${PIP_BIN} install numpy
+RUN ${PIP_BIN} install swig
+RUN ${PIP_BIN} install build
+RUN ${PIP_BIN} install wheel
+RUN ${PIP_BIN} install pyyaml
+RUN ${PIP_BIN} install cmake
+RUN ${PIP_BIN} install ninja
+RUN ${PIP_BIN} install -r /workspace/build/requirements.txt
+
+
+###
+#
+# Hardcode the cuda library path for the system loader
+
+RUN echo "/opt/nvidia/cuda/lib64" > /etc/ld.so.conf.d/cuda.conf
+RUN ldconfig -v
+
+
+###
+#
+# Build pytorch
+
+WORKDIR /workspace/build
+RUN ${PYTHON_VENV} -m build --wheel --sdist --no-isolation
+
+
+###
+#
+# Produce a clean image of build results for output from buildx
+
+FROM scratch
+COPY --from=bookworm_cuda /workspace/build/dist /

platform/Dockerfiles/pytorch/build.sh

@@ -0,0 +1,6 @@
+###
+#
+# Build pytorch and output the build results to  "${PWD}/target"
+
+mkdir -p "${PWD}/target"
+docker buildx build --progress plain --output type=local,dest="${PWD}/target" . -t pytorch:v2.3.1

platform/Dockerfiles/pytorch/pytorch-compute-86-override.patch

@@ -0,0 +1,20 @@
+diff --git a/aten/src/ATen/native/cuda/Blas.cpp b/aten/src/ATen/native/cuda/Blas.cpp
+index 8d3b3dbea7..5f04c0cdd1 100644
+--- a/aten/src/ATen/native/cuda/Blas.cpp
++++ b/aten/src/ATen/native/cuda/Blas.cpp
+@@ -820,7 +820,7 @@ static bool _scaled_mm_allowed_device() {
+     }
+     return false;
+ #else
+-    return dprops->major >= 9 || (dprops->major == 8 && dprops->minor == 9);
++    return (dprops->major == 8 && dprops->minor >= 0);
+ #endif
+ }
+
+diff --git a/third_party/cutlass b/third_party/cutlass
+index bbe579a9e3..56b46e2d13 160000
+--- a/third_party/cutlass
++++ b/third_party/cutlass
+@@ -1 +1 @@
+-Subproject commit bbe579a9e3beb6ea6626d9227ec32d0dae119a49
++Subproject commit 56b46e2d13875b46b8f6a03f9f5ac91e2bfdc01a